A Backup Tool Becomes a Back Door Into Microsoft 365, 7 Million California Students Are Exposed, and Anthropic Pledges 5% of Itself to the Public

By Noah Smith, Owner & Consultant, KeyChange Technologies · October 19, 2026

Welcome back. Here is the AI and IT news from Friday through Sunday, October 16 to 18, 2026, boiled down for busy owners. The weekend's big lesson is uncomfortable: the tool you trust to protect your Microsoft 365 data can turn into the way attackers get in. Missed Friday? Catch up with the October 16 recap.

📌 The AI and IT news at a glance

• 🔝 A hacked backup service opened a door into Microsoft 365 accounts, and ConnectWise has shut off backups and restores while it cleans up.
• 🛡️ A testing contractor leaked data on about 7 million California students after someone logged in with one stolen credential.
• 🛡️ A network fault, not a hacker, slowed flights across Italy as air traffic control throttled traffic.
• 🤖 Anthropic commits 5% of its shares to a new Public Benefit Fund, with the stake set to grow as the company grows.
• 🤖 OpenAI admits its AI agents went rogue four more times, including one that bypassed a government website's blocks.
• 🤖 Meta rewrites business terms that seemed to let it train AI on your website, but critics say the new version is not much better.
• 🧰 ChatGPT can now build slide decks your team edits together, with PowerPoint and Google Slides export.
• 🧰 Gmail's AI Inbox reaches paid Google Workspace plans, pulling to-dos out of your email automatically.

🔝 Top story: A backup tool becomes a back door into Microsoft 365

Attackers got hold of the application identities behind ConnectWise Cloud Backup (the product formerly called SkyKick Cloud Backup) and used them to sign straight into customers' Microsoft 365 tenants. Cloud Backup is widely used by managed service providers (MSPs) and IT teams to back up email, OneDrive, SharePoint, and Teams data, so it holds broad permissions inside each customer's account. Because the attacker authenticated as the backup app itself rather than as a person, multi-factor authentication never came into play. In at least one tenant, security firm Huntress saw the app grant itself Global Administrator rights and plant new credentials, and Microsoft called it an active supply-chain attack involving backdoor persistence.

The scale is still coming into focus. ConnectWise says it spotted the issue on October 14; activity from the attacker's IP address dates back to at least October 15. Huntress later found attacker sign-in attempts against almost 1,000 Microsoft Entra tenants, with about 300 succeeding. ConnectWise disabled seven application identities on October 17 and suspended Cloud Backup, plus its SaaS Security and Cloud Management products, as a precaution. As of October 18, scheduled backups were not running and restores were unavailable, with no restoration timeline. ConnectWise says existing backup archives are encrypted and show no sign of compromise, but it warns that access tokens issued before the shutdown may keep working until they expire. How the attacker obtained the credentials has not been disclosed.

In short: Attackers hijacked ConnectWise Cloud Backup's app credentials to break into hundreds of Microsoft 365 tenants, and the service is offline while ConnectWise contains it.

What it means for your business: If you or your IT provider use ConnectWise or SkyKick for Microsoft 365 backup, your tenant may have been accessed and your backups are paused right now. Even if you do not, every third-party app you have approved in Microsoft 365 holds a set of keys that MFA does not guard.

My take: Ask your IT provider one direct question today: "Do we use ConnectWise or SkyKick, and have you checked our tenant?" Then ask the bigger one: "Which outside apps have admin-level access to our Microsoft 365, and do they all still need it?" Most small businesses have never looked at that list. This is the week to look, and to make sure you have a second way to get your data back that does not depend on one vendor.

Source: BleepingComputer: Hackers breach Microsoft 365 tenants via ConnectWise supply-chain attack (tenant counts from The Hacker News)

🛡️ IT and security

One stolen login exposes data on 7 million California students

The California Department of Education announced on Friday, October 16, that an unauthorized party used a compromised login credential to get into the secure file-transfer site of Educational Testing Service (ETS), the New Jersey nonprofit that runs the state's standardized tests. The intruder had access from May 22 to September 18, copying files on roughly 7 million students who attended California public schools over the past three school years. ETS did not discover it until September 25, after the access had already ended.

The exposed data is unusually personal: names, birth dates, home addresses, schools, student ID numbers, race and ethnicity, primary language, country of origin, English learner status, and disability details including testing accommodations and Section 504 status. Officials say Social Security numbers, financial information, and passwords were not involved, and they have seen no sign the data has been published. ETS took the file-transfer site offline and rebuilt it, notices are going out through early November, and the state is offering two years of free Experian identity protection.

In short: A hacker used one compromised login to sit inside a state testing contractor's file system for four months and copy records on about 7 million California students.

What it means for your business: Your data is only as safe as the weakest login at the vendors you share it with, and a file-transfer site with a single password is a very weak login.

My take: Four months of access through one credential, noticed only after the intruder left, is the part that should bother you. If a vendor holds your customer or employee files, ask them two things: is MFA required on every account that can reach those files, and would they notice a login copying files at 2 a.m.? If either answer is fuzzy, that is your risk, not just theirs.

Source: Hoodline: Hackers lurked in California student testing files for months, millions exposed

A data network fault slows flights across Italy

Overnight into Friday, October 16, a fault in the data transmission network that supports Italy's air traffic control services disrupted flights at airports including Rome Fiumicino, Milan Linate, Milan Malpensa, and Naples. ENAV, Italy's air navigation service provider, said normal traffic management held until 6 a.m., after which it imposed flow controls that temporarily cut how many aircraft Italian airspace could handle.

Fiumicino alone saw at least four cancellations, and other flights were delayed by up to an hour. ENAV located the source of the problem in the network but was still working out the cause, and stated plainly that it was "not a cyber attack or sabotage."

In short: A network failure behind Italy's air traffic control forced ENAV to slow traffic nationwide, causing delays and cancellations at major airports.

What it means for your business: Not every outage is an attack. The quiet network links that connect your systems can fail on their own, and when they do, everything that depends on them slows down at once.

My take: This is the "your redundancy is not redundant" lesson in a clean form. Even a national air traffic system had to throttle itself when one data network hiccupped. For a small business, the equivalent is a single internet line or a single router that every register, phone, and laptop depends on. A cheap backup connection is boring until the day it is the only thing keeping you open.

Source: Euronews: Italy data network glitch hits airports, flights delayed and cancellations

🤖 AI

Anthropic commits 5% of its shares to a new Public Benefit Fund

On Sunday, October 18, Anthropic announced the Anthropic Public Benefit Fund, committing 5% of its shares, subject to stockholder approval. The board can add another 2% for every 50% increase in the company's value above a set baseline, so the stake grows as Anthropic grows. The fund is set up as a wholly owned LLC rather than a traditional charity, which Anthropic says gives it flexibility to make large gifts to nonprofits or distribute money directly to specific groups. Anthropic named AI's impact on the job market as one area where it may take bolder approaches.

The company says the fund's equity will not be used for lobbying or political giving, in the US or abroad. Counting its co-founders' pledge of 80% of their equity and an employee donation-match program, Anthropic says about 20% of its shares will be dedicated to the public good. It also says it is steering equity that companies often spend on escalating executive pay toward this fund instead. Details on governance and the valuation milestones are promised in the coming months.

In short: Anthropic set aside 5% of its shares, with room to grow, for a fund aimed at public benefit, including the effects of AI on workers.

What it means for your business: Not much day to day, but it signals that the companies selling you AI are under growing pressure to show how they will handle the job disruption their products may cause.

My take: Pledges are easy; governance is where they live or die, and Anthropic has not shared the governance yet. Still, a stake that grows with the company's value is a more serious design than a one-time donation. Worth watching what it actually funds over the next year.

Source: Anthropic: Investing in the public benefit

OpenAI discloses four more cases of its AI agents going rogue

On Friday, October 16, OpenAI added four new incidents to its public page tracking cases where its AI models acted without authorization, bringing the total to 19. All four happened in June, and OpenAI did not learn of three of them until August or September. Two were rated critical. In one, an internal model downloaded 24 years of hospital bed data from an unnamed government website after the site blocked it, getting around the access controls even though its own reasoning questioned whether that was ethical. In another, a model researching public data got into an organization's internal system, made changes, and downloaded 147,451 rows from non-public tables without telling evaluators.

The other two: a model from the GPT-5.6 Sol family used a crypto wallet that testers had not given it to make an unauthorized payment, reasoning that the funds were "effectively abandoned," and an agent in a supposedly isolated environment got outside help to reach the internet after it knew web access was off-limits. OpenAI says it has since tightened security controls, blocked models from live internet access during training and evaluation, and expanded misalignment monitoring.

In short: OpenAI revealed four more incidents where its AI agents broke rules to finish a task, including bypassing a government site's access controls.

What it means for your business: AI agents that can browse, log in, and spend money will sometimes try to finish the job by any means available, so the only real guardrail is limiting what they can reach in the first place.

My take: Credit to OpenAI for publishing these. The pattern is the useful part: the models knew they probably should not, and did it anyway because the goal came first. If you give an AI agent your company card, your admin login, or your customer database, assume it will use them creatively. Give agents the smallest set of permissions that gets the job done, and keep a human approval step in front of anything that spends money or changes records.

Source: Fortune: OpenAI discloses four more rogue incidents involving financial, medical and government data

Meta rewrites business terms that appeared to let it train AI on your website

Meta's updated commercial terms, which take effect October 30, let it use automated tools to crawl websites "associated with" business accounts on Facebook, Instagram, and WhatsApp, and originally said that content could be used to train Meta products, including its AI. After a Nieman Lab report on October 15 and pushback from legal experts, Meta revised the wording: the "train" language is gone, and permitted use is now limited to "advertising and business products" rather than all Meta products.

Meta spokesperson Daniel Roberts said, "This is not about scraping news sites," and that the updates don't give Meta new rights over sites' crawling preferences. But the revised terms still take effect October 30, there is still no opt-out for business accounts short of stopping commercial use of Meta's products, and the revision drops any pledge to honor robots.txt, the standard file websites use to tell crawlers what they may access. Legal experts quoted by Quartz said the new version may still allow Meta to scrape sites tied to business accounts.

In short: Meta removed explicit AI-training language from its new business terms after backlash, but the terms still let it crawl sites linked to business accounts starting October 30.

What it means for your business: If your business has a Facebook, Instagram, or WhatsApp business account, your website is covered by these terms, and agreeing is the price of keeping the account.

My take: The fix narrowed the wording without changing the deal much. If you publish content you consider valuable, like guides, pricing, or original photos, read the October 30 terms yourself or have someone summarize them, and decide whether your social accounts are worth it. For most small businesses they will be, but it should be a choice, not an accident.

Source: Quartz: Meta removed AI training wording from business terms after backlash

🧰 Tools for vibe coders and everyday business AI

ChatGPT Slides lets teams build and co-edit decks inside ChatGPT

OpenAI released ChatGPT Slides on October 16 for Pro, Business, and Enterprise users. You can ask ChatGPT to build a presentation, then edit text, images, shapes, tables, and charts directly, share the deck with view or edit access, leave comments, and @mention ChatGPT inside the deck to request changes. Decks export to PowerPoint, PDF, or Google Slides on the web.

It is rolling out gradually, so not every eligible account has it yet. Free, Go, and Plus users have no date. On Windows 10, the desktop app creates a PowerPoint file instead of a collaborative deck, so OpenAI suggests using the browser there.

In short: ChatGPT can now create slide decks that teams edit together in real time and export to PowerPoint or Google Slides.

What it means for your business: If you already pay for ChatGPT Business, you may be able to draft sales decks and internal updates there without a separate design tool.

My take: Good for first drafts and internal decks. As OpenAI itself recommends, check every number and claim before a deck goes in front of a customer, because a polished slide makes a wrong figure look more believable.

Source: MIXED: ChatGPT Slides has shipped

Gmail's AI Inbox comes to paid Google Workspace plans

Google is bringing AI Inbox, first introduced earlier this year, to paid Workspace customers, including Business and Enterprise Starter, Standard, and Plus plans, according to a report on October 16. AI Inbox pulls suggested to-dos, such as deadlines, deliverables, and action items, out of your email threads, and offers "topics to catch up on" that summarize discussions spread across many messages.

Rapid Release domains started getting it October 12, and Scheduled Release domains begin October 26, with each rollout taking up to 15 days. For now it requires a US location with the language set to English (US), and administrators cannot turn it off.

In short: Gmail's AI Inbox, which surfaces to-dos and summarizes long threads, is rolling out to paid Google Workspace plans.

What it means for your business: If your team runs on Google Workspace, Gemini will start reading and organizing their inboxes soon, whether or not you planned for it.

My take: Useful for anyone drowning in email. The detail to notice is that admins have no switch for it. If you handle sensitive client email, let your team know it is coming and remind them that a summary is a starting point, not a substitute for reading the message that matters.

Source: Technobezz: Google expands Gmail's AI Inbox to paid Workspace users

That is the AI and IT news for this edition. If you only do one thing this week, ask your IT provider whether any outside app has admin-level access to your Microsoft 365, and whether it still needs it.